System Identification without Lennart

Ljung: what would have been different?

Michel Gevers
To Lennart, for a friendship of more than thirty years,
that has never faltered.
Abstract:

This chapter presents a personal view on the development

of identification theory in the control community, starting from the year
1965. We show how two landmark papers, (Ho and Kalman, 1965) and
(strm and Bohlin, 1965), gave birth to two main streams of research
that have dominated the development of system identification over the last
fourty years. The Ho-Kalman paper gave a first solution to state-space
realization theory, which led to stochastic realization, and much later to
subspace identification. The strm-Bohlin paper laid the foundations
for the highly successful Prediction Error methods based on parametric
input-output models. The chapter highlights the key influence of Lennart
Ljung on the development of Prediction Error Identification; it shows how
his seminal contributions have profoundly changed the communitys view
on identification from a search for the elusive "true system" to a goaloriented design problem.

Introduction

The development of identification theory in the control literature followed on the

heels of the development of model-based control design around 1960. Up until
the late 1950s, much of control design relied on Bode, Nyquist and Nichols
charts, or on step response analyses. These techniques were limited to control
design for single-input single-output (SISO) systems. Around 1960, Kalman
introduced the state-space representation and laid the foundations for state-space
based optimal filtering and optimal control theory, with Linear Quadratic (LQ)
optimal control as a cornerstone for model-based control design.
The availability of these model-based control design techniques put pressure
on the scientific community to extend the fields of application of modern control design beyond the realm of mechanical, electrical and aerospace applications, for which reliable models were easily available. Thus the need arose to
develop data-based techniques that would allow one to develop dynamical models for such diverse fields as process control, environmental systems, biological
and biomedical systems, transportation systems, etc.

System Identification without Lennart Ljung: what would have been different?

Much of the early work on identification was developed by the statistics,

econometrics and time series communities. Even though the statistical theory of
parameter estimation has its roots in the work of Gauss (1809) and Fisher (1912),
most of the theory of stationary stochastic processes was developed during the
period 1920 to 1970. We shall not describe this work here, because we want to
focus on the engineering views and developments of system identification. An
excellent review of the history of system identification and time series analysis
in the statistics community can be found in Deistler (2002).
Although a lot of results had already been established in the statistics and
econometrics literature, one can view 1965 as the birthyear for identification theory in the control community, with the publication of two seminal papers, Ho
and Kalman (1965) and strm and Bohlin (1965). These two papers paved the
way for the development of the two mainstream identification techniques that still
dominate the field today: subspace identification and prediction error identification. The former is based on projection techniques in Euclidean space, the latter
is based on the minimization of a parameter dependent criterion.
The Ho-Kalman paper provided the first solution to the determination of a
minimal state-space representation from impulse response data. The solution
of this deterministic realization problem was later extended by Akaike (1974)
and others to stochastic realization, where a Markovian model is obtained for a
purely random process on the basis of covariance data. This technology, based on
canonical correlation analysis, was extended in the early nineties to processes that
also contain a measurable (control) input, and was then rebaptized as subspace
state-space identification.
The strm-Bohlin paper introduced into the control community the Maximum Likelihood framework that had been developed by time series analysts
for the estimation of the parameters of difference equation models. These were
known in the statistical literature by such esoteric names as ARMA (AutoRegressive Moving Average) or ARMAX model (AutoRegressive Moving Average with
eXogeneous inputs). These models, and the Maximum Likelihood framework,
were there to stay, since they gave rise to the immensely successful Prediction
Error Identification framework.
In 1970, Box and Jenkins published their book Time series analysis, forecasting and control, Box and Jenkins (1970), which gave a major impetus to
applications of identification. Indeed, the book gave a rather complete recipe for
identification, all the way from initial data analysis to the estimation of a model.
In the spirit of the time series analysis methods of the time, it relied on correlation
analysis for the determination of model structure. For about 15 years, it remained
the major high quality reference book on system identification. Other important
references of the time were the survey paper strm and Eykhoff (1971) and
the special issue on system identification and time series analysis published by
the IEEE Transactions on Automatic Control in December 1974. The strm

Michel Gevers

and Eykhoff survey was to be used by many young researchers of the time as
a stepping stone for future work. It explained the state of the art as much as it
displayed some of the important open questions of the time. One of these was
the identification of closed-loop systems, for which the Hankel-based projection
methods (based on cross-correlation information) had been shown to fail.
From about the mid-seventies, the Prediction Error (PE) framework completely dominated identification theory and, perhaps more importantly, identification applications. Just about all of the activity at that time was focused on
the search for the true system, i.e. it dealt with questions of identifiability,
convergence to the true parameters, and asymptotic normality of the estimated
parameters. Much of that activity dealt with identifiability problems for multivariable systems and for closed-loop systems.
Around 1976 the first attempts were made to view system identification as an
approximation problem, in which one searches for the best possible approximation of the true system within some model class: Ljung (1976); Anderson et al.
(1978); Ljung and Caines (1979). The prevailing view changed consequently
from a search for the true system to a search for and characterization of the
best approximation. Hence, the characterization of the model errors (bias error
and variance error) became the focal point of research. For control engineers, the
object of primary interest is the model, in particular the transfer function model,
rather than the parameters which are just a vehicle for the description of this
model. As it turns out, the research on bias and variance error moved remarkably
swiftly from the characterization of parameter errors to that of transfer function
errors, thanks to some remarkable analysis of Ljung based on the idea of letting
the model order go to infinity: Ljung (1985); Wahlberg and Ljung (1986).
The work on bias and variance analysis of identified models of the eighties
then led, almost naturally, to a new perspective in which identification became
viewed as a design problem. With an understanding of the effect of the experimental conditions, the choice of model structure, and the choice of criterion on
the quality of the identified model, one can tune these design variables towards
the objective for which the model is being identified: Gevers and Ljung (1986).
The book System identification: Theory for the user, Ljung (1987), has had a
profound impact on the engineering community of system identifiers. It squarely
put forward the view of system identification as a design problem, in which the
model use plays a central role. This viewpoint clearly distinguishes the field from
the statistical literature on system identification and time series analysis, where
the prevailing view is that the model must explain the data as best as possible.
The observation that the model quality can be tuned, through the choice of
appropriate design variables, towards the eventual objective for which the model
is being built opened the way to a flood of new activity that took place in the
nineties and continues up to this day. The major application of this new paradigm
is the situation where a model is built with the view of designing a model-based

System Identification without Lennart Ljung: what would have been different?

controller. Thus, identification for control has blossomed, since around 1990.
Because that topic embraces many aspects of identification and robust control
theory, it has also opened or reopened new research interest in areas such as
experiment design, closed-loop identification, frequency domain identification,
uncertainty estimation, and data-based robust control analysis and design.
The present chapter attempts to exhibit both the continuity and the motivation for the developments that took place in system identification in the last fourty
years, and also the significant new departures and insights that came as the result of some important breakthroughs. In doing so, this chapter will show how
Lennart Ljung was responsible for several of these breakthroughs.

The milestone papers

2.1

Deterministic realization theory

In 1965, Ho and Kalman (1965) provided a first solution to a challenging system

theoretical problem that became known as the state-space realization problem. It
can be stated as follows.
Construct a minimal state-space realization

xt+1 = Axt + But
yt
= Cxt
for an input-output model described by its impulse response matrices (also called
Markov parameters) Hk Rpm :
yt =

Hk utk .

k=1

The problem is thus to replace the infinite description

H(z) =

Hk z k

k=1

by a finite description with A Rnn , B Rnm , C Rpn , such that

H(z) = C(zI A)1 B
with dim(A) minimal. This problem can be split up into two parts: (i) find
the McMillan degree of H(z), which is then the minimal dimension of A; (ii)
compute the matrices A, B, C. The key tool for the solution of this problem is the

Michel Gevers

Hankel matrix, and its factorization into the product of an infinite observability
matrix times an infinite controllability matrix:

C
H1 H 2 H3 H 4 . . .
H2 H3 H4 H5 . . . CA 


H = H3 H4 H5 H6 . . . = CA2 B AB A2 B . . .

..
..
..
..
..
.
.
.
.
.
If the McMillan degree of H(z) is n, then
1. rank H = n
2. A Rnn , B Rnm , C Rpn such that Hk = CAk1 B.
It took years of research to go from the theoretical results described in Ho and
Kalman (1965) to a numerically reliable realization algorithm. However, all the
key insights were present in the 1965 paper, and they were to have a profound
impact on linear system theory, and on realization and identification theory.

2.2

The Maximum Likelihood framework

In complete contrast to the state-space formulation of Ho and Kalman, the landmark paper strm and Bohlin (1965) introduced the Maximum Likelihood method for estimating the parameters of input-output models in ARMAX form:
A(z 1 )yt = B(z 1 )ut + C(z 1 )et
where {et } is a sequence of independent identically distributed Normal (0, 1)
random variables. The Maximum Likelihood (ML) method was well known and
had been widely studied in mathematical statistics and time series analysis. However, what is remarkable about the strm-Bohlin paper is that the authors not
only gave a complete algorithmic derivation of the ML identification method for
ARMAX models, but also presented all analysis results that were available at that
time, such as the consistency, asymptotic efficiency and asymptotic normality of
the parameter estimates, the persistence of excitation conditions on the input signal in connection with the order of the model, the model order validation on the
basis of the whiteness of the residuals, etc.
The concepts and notations introduced by strm and Bohlin in 1965 have
been with us for almost 40 years now. Indeed, the following household notations
of the identification community can all be found in this milestone paper:
the residuals C(z 1 )t = A(z 1 )yt B(z 1 )ut
PN
the cost criterion V () = 21 t=1 2t

System Identification without Lennart Ljung: what would have been different?

the parameter estimate = arg min V ()

2 =
the white noise variance estimate

2
NV

().

The publication of strm and Bohlin (1965) gave rise to a flurry of activity
in parametric identification. It also established the basis for the adoption of the
Prediction Error framework. The step from Maximum Likelihood to Prediction
Error essentially consists of observing that, under an assumption of white Gaussian noise in the ARMAX model, the maximization of the likelihood function of
the observations is equivalent to the minimization of the sum of the prediction
errors. The Prediction Error framework then consists of adopting the minimization of a norm of the prediction errors as a reasonable criterion for parameter
estimation, even in the absence of any known probability distribution for the observations. Such suggestion had already been made by Mr. Gauss himself, Gauss
(1809), as observed in the fascinating paper strm (1980).

From deterministic to stochastic realization

The combination of the deterministic realization theory based on the factorization

of the Hankel matrix, and of the theory of Markovian and innovations representations gave rise to the stochastic theory of minimal realizations. The stochastic
realization problem can be stated as follows.
Given the covariance sequence {Rk , k = 1, 2, . . . , } of a zero-mean stochasT
tic process {yt }, where Rk , E{yt ytk
}, find a minimal Markovian representation for the process {yt }, of the form

xt+1 = Axt + Gwt
(1)
yt
= Cxt + vt


wt
where
is a zero-mean white noise sequence with covariance matrix
vt
(

T ) 

Q
S
wt
wt
W =E
=
.
vt
vt
ST R
This problem amounts to finding state-space matrices {A, G, C} with n = dim(A)
minimal, and the elements Q, S, R of the covariance matrix W such that the covariance of the output of (1) is exactly Rk .
Observe that the covariance of the output of the Markovian representation (1)
is given by Rk = CAk1 N with N = AC T + GS for k > 0, and R0 =
CC T + R, where is the state covariance: , E{xt xTt }.

Michel Gevers

The stochastic realization problem was studied very intensively during the
early seventies in connection with innovations theory and spectral factorization
theory: Akaike (1974); Gevers and Kailath (1973); Faurre (1976). The first step
of the solution consists of observing that the Hankel matrix made up of the covariance sequence can be factored as

C
R 1 R2 R 3 . . .
R2 R3 R4 . . . CA 


H = R3 R4 R5 . . . = CA2 N AN A2 N A3 N . . .

..
..
..
..
.
.
.
.
where Rk = CAk1 N with N defined as above. The Ho-Kalman algorithm
mentioned above allows one to determine the minimal dimension n as the rank
of H, and to compute the matrices C, A and N from the factorization of H.
There are various ways of performing the second step, which consists of computing the missing elements G, Q, R, S of the Markovian representation (1) from
C, A, N and the output variance R0 . One way is to use a particular version of
the Markovian representation called an innovations model:

t+1 = At + Kt
(2)
yt
= Ct + t
where {t } is a white noise sequence with covariance matrix = E{t Tt }.
= E{t tT } the covariance of the state of this model, and imposing
Denoting
that the covariance sequence of the output of this innovations model (2) is {Rk },
K and :
one gets the following three constraints on the unknown quantities ,

R0

T + KK T
= AA

= AC T + K
T +
= C C

The first (Lyapunov) equation follows directly from the state equation of the
Markovian model, while the other two constraints are imposed by the matching of the output covariance Rk : see above. The solution of this second step is
obtained via an associated Riccati equation.
An interesting aspect of the solution of the stochastic realization problem
given by Akaike was the definition of the state of the innovations model as the
set of canonical correlations obtained by projecting the space of future outputs
onto the space of past outputs Akaike (1975). This insight formed the basis for
the later work on subspace identification; it also gave rise to extensive studies of
the interface between past and future observation spaces, called minimal splitting
subspaces: see e.g. Katayama and Picci (1999).

System Identification without Lennart Ljung: what would have been different?

Another important outcome of the stochastic realization and innovations theories of the seventies was the covariance equivalence established between the
Markovian realization (1) and its innovations realization (2). In particular, this
means that, even if a Markovian model (1) has been constructed as a physical
model for yt , driven by two independent white noise sources wt and vt (i.e.
S = 0, say), it can be rewritten as the Markovian innovations model (2) driven
by a unique white noise source t with the same output covariance {Rk }. The
same equivalence applies to models with deterministic inputs. This means that
any (physical) state-space model

xt+1 = Axt + But + Gwt
yt
= Cxt + vt
can be rewritten as an innovations model for yt driven by a unique noise source
t :

t+1 = At + But + Kt
yt
= Ct + t .
The input-output equation of this state space innovations model is:
yt

= C(zI A)1 But + [C(zI A)1 K + I]t

= G(z)ut + H(z)t

where
G(z) = C(zI A)1 B, H(z) = I + C(zI A)1 K.
Observe that the transfer functions G(z) and H(z) have the same poles, and that
H(z) is monic. Thus, this model is equivalent to an ARMAX model:
A(z 1 )yk = B(z 1 )uk + C(z 1 )k .
This theory established a nice link between Markovian models obtained from
first principles modeling, their corresponding state-space innovations models,
and the ARMAX input-output models used in Maximum Likelihood (ML) or
Prediction Error (PE) identification. It also gave a solid theoretical justification
for the use of ARMAX models to represent stationary linear Markov processes.

4
4.1

The golden years: 1975-1985

The big cleanup

The years 1975 to 1985 saw a frantic activity in system identification in the engineering community. The parametric input/ouput methods, in large part because
increased computer speed and the development of special purpose identification

Michel Gevers

software made it more and more feasible to perform the iterative minimization
of a cost criterion over a range of possible model structures. This is the period
where many authors were putting their name on a new combination of model
structure and method, with the inevitable claim about the supremacy of their new
combination over existing methods. New methods were pouring out constantly
in the scientific journals.
Some solid cleaning work was required, and it was one of L. Ljungs major
contributions to perform this cleaning work. It consisted of clearly separating
two independent concepts: the choice of a parametric model structure, which
was seen as just a vehicle for computing predictions and hence parameter dependent prediction errors; and the choice of an identification criterion, which was to
be chosen as some positive function of these prediction errors and hence of the
parameter vector: see e.g. Ljung (1978). All existing parametric identification
methods could then be seen as particular cases of this prediction error framework.
Thus Ljung introduced the generic input-output model structure
yt = G(z, )ut + H(z, )et

(3)

with G(z, ) and H(z, ) parametrized as rational transfer functions, and with
et white noise. All commonly used model structures, such as ARX, ARMAX,
OE, BJ, FIR, were special cases of the structure (3). From such model structure,
one can derive the parameter-dependent one-step ahead prediction yt|t1 (), and
hence the one-step ahead prediction error t () = yt yt|t1 (). From a set of
N data, Z N , and hence of N prediction errors, {t (), t = 1, . . . , N }, one can
then define a criterion
VN (, Z N ) =

N
1 X
l(t ()),
N t=1

(4)

where l(.) is a positive scalar-valued function. The minimization of VN (, Z N )

with respect to in some domain D then yields the parameter estimate:
N = arg min VN (, Z N ).
D

(5)

The methodological work of Ljung culminated with the publication of his

book, Ljung (1987), which has become the standard reference book in system
identification. The impact of his book was greatly amplified by the simultaneous production by L. Ljung of the MATLAB identification toolbox, whose tools
were directly connected to the theory developed in the book. Ljungs book was
complemented by that of Stoica and Sderstrm, who adopted the same clear
distinction between choice of model structure and choice of criterion; their book
focused less on design issues, but more on analysis and on alternative criteria, in
particular criteria based on correlation methods and instrumental variables: see
Sderstrm and Stoica (1989).

10

4.2

System Identification without Lennart Ljung: what would have been different?

Breakthroughs for MIMO and for closed-loop systems

During that same period, some important theoretical breakthroughs were made in
two directions. The first was the elucidation of the manifold structure of multiinput multi-output systems. The second consisted of a range of identifiability
results for linear systems identified under closed-loop conditions.
Many authors contributed to the solution of both problems, and we cannot possibly name them all. Let us just mention that the manifold structure of
MIMO systems, together with the key role of the Kronecker (or structure) indices, was elucidated in Hazewinkel and Kalman (1976) and Clark (1976). Subsequently, many authors worked on methods for the estimation of these structure indices; others studied the relationship between the canonical (or pseudocanonical) forms in state-space and in ARMA form: Deistler and Gevers (1981);
Van Overbeek and Ljung (1982); Picci (1982); Rissanen (1982); Wertz et al.
(1982).
As for the identifiability of closed-loop systems, one of the earliest solutions
was provided by the famous Swedish trio made up of Gustavsson, Ljung and
Sderstrm1 , all PhD students of K.J. strm at the time: see Sderstrm et al.
(1976). They showed, among other things, that in many situations of practical
interest, the direct application of a PE identification method to input-output data
allows one to identify the open-loop plant, despite the presence of a feedback
controller. Other closed-loop identifiability results covered indirect methods (in
which the closed-loop transfer function is identified first, and the plant model is
then derived from it using knowledge of the controller) and the so-called joint
input-output method: see Ng et al. (1977); Gustavsson et al. (1977); Sin and
Goodwin (1980); Anderson and Gevers (1982).

4.3

System identification viewed as approximation

For most of the sixties and seventies, the prevailing assumption was that the system was in the model set: S M. Thus, the focus of research was on questions
of convergence to the true system and of statistical efficiency of the parameter
estimates. In the mid-seventies, the first attempts were made to view system
identification in the context of approximation: Ljung (1976); Anderson et al.
(1978); Ljung and Caines (1979). This marked the beginning of an entirely new
era, in which the elusive search for a linear time-invariant true system was progressively abandoned to give way to the search for a best approximate model
within some a priori chosen model set M. With the idea of model approximation
came of course the idea of model error, and hence the desire to characterize this
model error.
1 Over a 4-year period they jointly published no less than 6 important papers on various aspects of
system identification.

11

Michel Gevers

4.4

The birth of

In statistics, the natural way to analyse estimation errors is through the concepts
of bias and variance error of the parameter estimate. However, in the context of
model sets that do not contain the true system, the concept of parameter error
becomes meaningless; since there are no true parameters, the bias error in the
parameter estimate does not make sense. The object of interest is the transfer
function, not the parameters that are used to represent it. By observing that, under
reasonable conditions, the parameter estimate N converges to a well-defined ,
lim N = = arg min lim EVN (),

D N

Ljung introduced the concept of bias error and variance error of the transfer function estimate by defining the following decomposition of the total transfer function error, at a given frequency :
G0 (ej ) G(ej , N )

= G0 (ej ) G(ej , )
{z
}
|
bias error

+ G(ej , ) G(ej , N )
{z
}
|
variance error

Within this framework, he and his collaborators derived approximate asymptotic

expressions for the transfer function variance, and integral expressions for the
transfer function bias that were to become the cornerstone of the next major phase
of the development of system identification: the design phase; see Ljung (1985);
Wahlberg and Ljung (1986).

Identification as a design problem

If identification is viewed as approximation, then one knows and accepts that the
model is erroneous. If the model is to be used for a specific purpose (as is most
often the case), then perhaps one can construct a model in such a way that the
model errors do not penalize too much the goal for which the model is being
built. This is the whole idea behind goal-oriented identification. More precisely,
if one can understand the connection between
identification design (experiment design, choice of model structure, choice
of criterion) and model quality on the one hand,
the effect of model quality on the intended model application on the other
hand,

12

System Identification without Lennart Ljung: what would have been different?

then one can formulate the identification problem as a goal-oriented design problem.
The first few steps of this new paradigm were laid in Gevers and Ljung
(1986); Wahlberg and Ljung (1986); Ljung (1987). This new (engineering) way
of looking at the identification problem opened up a vast new window of opportunities for research. In particular, this viewpoint was instrumental in the development, from around 1990, of a field that has seen an enormous activity ever since,
both on the theoretical front and in practical applications, namely identification
for control.

System identification in the nineties

At the triennial IFAC Symposium on System Identification held in Budapest in

1991, there was a feeling among many participants that most of the important
problems in system identification had been solved and that the golden age of
identification was over. This prediction proved to be wrong. The research on
system identification was pulled all through the nineties essentially by two catalysts, whose first feeble signs emerged around 1990: subspace-based identification, and identification for control. In addition, important new research activity
took place in frequency domain identification, in closed-loop identification, and
in the development of a range of new methods for the quantification of model
uncertainty.

6.1

Subspace-based identification

The reasons for the emergence of subspace identification are almost certainly
to be found in the state of the art of identification of multivariable systems in
the eighties. Even though the manifold structure of MIMO (multi-input multioutput) systems had been completely characterized in the late seventies, the practical problem of identifying MIMO systems remained wide open. Indeed, the
estimation of the structure indices that characterize the parametrizations of multivariable systems remained very tricky, and led to ill-conditioned numerical procedures. Thus, there was a great incentive for the development of simple, albeit
suboptimal procedures, based on the numerically robust Singular Value Decomposition and Least Squares techniques, that completely bypass the need for the
estimation of structure indices. The development of subspace-based identification methods filled a much needed gap, because in that framework the handling
of MIMO systems causes no additional difficulty.
A major difficulty was that the projection methods developed by Akaike,
which were based on canonical correlation analysis, were not easily extendable
to output data that contained, besides the stochastic components, also a contri-

Michel Gevers

13

bution due to a measured input. In the early nineties, several research teams
managed to crack this nut; they provided several closely related solutions to the
problem: see Larimore (1990); Van Overschee and De Moor (1994); Verhaegen
(1994); Viberg (1995). These first few solutions opened the way for a lot of research on the properties of subspace-based identification, their connection with
stochastic realization theory, and on improvements of the numerical procedures:
see e.g. Chui and Maciejowski (1996); Katayama and Picci (1999); Chiuso and
Picci (2002).

6.2

Identification for control

Identification for control has been the major outlet for the new paradigm of system identification as a design problem. The reasons for this are many: (i) in the
systems and control community of system identification, control is very often the
main motivation for model building; (ii) it has been observed that high performance control can often be achieved with very simple models, provided some
basic dynamical features of the system are accurately reflected; (iii) a powerful
robust control theory, based on models and uncertainty sets, had been developed all through the eighties but the models and uncertainty sets used were not
data-based for lack of a proper theory; (iv) the identification for control research
delivered iterative model and controller tuning tools that were intuitive, practical
and easy to implement by the process engineers.
Whereas the building blocks for goal-oriented identification were laid around
1986, the first specific contributions in which identification and control design
were looked upon as a combined design problem appeared only around 1990.
The plenary Gevers (1991) at the 1991 IFAC Symposium on System Identification addressed many of the key issues about the interplay between the identification of a reduced order model and the design of a controller from such model;
however, it was more an agenda for research than a presentation of solutions.
Indeed, in 1990 there was very little understanding about the interplay between
system identification and robust control. The two theories had been developed
by two separate communities that had had very little contact with one another.
The robust control community had developed a robust analysis and design
theory that was based on uncertainty descriptions which were not based on data
but on prior assumptions. The identification community had delivered bias and
variance error descriptions that were not explicit, and that were certainly not
transferable at the time to the toolboxes used in robust control analysis and design. More importantly, neither community had given much attention to the interaction between model building and control design: what are the qualities that
a model must possess (or, conversely, what are the plant-model errors that are
acceptable) if the model is to be used for the design of a controller that must
achieve a given level of performance on the plant?

14

System Identification without Lennart Ljung: what would have been different?

In July 1992, the IEEE Transactions on Automatic Control devoted a special

issue to system identification for robust control design. In retrospect, and in
keeping with the observation just made, such issue was perhaps premature, given
the paucity of results that were available at that time. About half of the papers
in that special issue did not really deal with identification for control, but with
the estimation of uncertainty sets, without any account taken of control-oriented
design issues. A few papers in that special issue did address the joint identification and control design paradigm, notably Bayard et al. (1992) and Schrama
(1992). They produced one of the first key results in identification for control:
the necessity of an iterative scheme for the design of a control-oriented nominal
model; this observation had in fact been first made in Liu and Skelton (1990).
The first half of the nineties produced a string of results on the design of
control-oriented nominal models. These results were produced by different teams
who used their favorite combinations of identification criteria and control design criteria: Schrama and Bosgra (1993); Lee et al. (1993); strm and Nilsson
(1994); Zang et al. (1995); de Callafon and Van den Hof (1997). They confirmed
the necessity of using an iterative scheme of model updates and controller updates, and they produced significant evidence - but no hard proofs - about the
advantages of performing the identification in closed loop, rather than in open
loop, when the model is to be used for the design of a new controller. Of course,
the closed-loop experimental conditions that produce a specific input signal spectrum can always be mimicked by an open loop experiment with the same input
spectrum, but the advantage of the closed-loop experiment is that the appropriate
input spectrum is automatically generated by the unknown system itself.
The iterative schemes of identification and control design had a remarkably
fast transfer into the world of applications. There were two main reasons for this:
whereas the industrial world was still living with the belief that one should
open the loop to perform a valid identification experiment, here was a
new theory that showed the benefits of closed-loop identification; this came
as welcome news to process control engineers who had never really liked
the idea of opening the loop;
in the process industry, thousands of measurements are flowing into the
computer; here was a theory that showed how these data could be used for
the design of a better controller.
The early work on identification for control focused on control-oriented identification criteria, i.e. on obtaining a nominal model whose bias error distribution
was tuned for control design. This means that the nominal control performance
obtained with the optimal controller computed from the nominal model is close
to the actual control performance obtained with the same controller on the actual
plant.

Michel Gevers

15

More recently, attention has shifted to the distribution of the variance error
of the identified models, i.e. to the estimation of control-oriented uncertainty
sets: Kosut and Anderson (1994); Mkil et al. (1995); de Vries and Van den
Hof (1995); Bombois et al. (2000). The idea is that, since one can manipulate the
shape of the model uncertainty set by the choice of the experimental conditions
under which the new model is identified, one should attempt to obtain a model
uncertainty set that is tuned for control design. Even though many new insights
have been gained on the interplay between uncertainty sets estimated from data
and corresponding sets of stabilizing controllers, there is at this point no clear
view as to the most operational definition of a control-oriented uncertainty set.
One view is that the corresponding set of controllers achieving stability and the
required performance with all models of that set should be as large as possible
Gevers et al. (2003). Another view is that the worst-case performance achieved
by some optimal robust controller with all models in this uncertainty set should
be as close as possible to the performance achieved with the central (nominal)
model.
The work on identification for control has had many beneficial side effects: it
has forced the scientific community to reassess some truths that had been considered to be firmly establised and to reopen some research questions that had
been considered to be completely settled. Identification for control has triggered
an enormous new research activity on the estimation of data-based uncertainty
sets for identified models, as well as on the identification of systems operating in
closed loop. An intense debate has been reopened on the relative merits of model
or controller reduction versus the direct identification of a control-oriented restricted complexity model, leading to the notion of near-optimal restricted complexity model introduced in Hjalmarsson (2005), a must-read for any researcher
in identification for control. Finally, identification for control has led to the recent rebirth of the topic of optimal experiment design, a subject that had been
very active in the 1970s but which had been almost completely abandoned since
then. The overview paper Gevers (2005) presents up-to-date results on the use of
optimal experiment design in the context of identification for control, a topic that
is also very present in Hjalmarsson (2005).
In the remainder of this chapter, we briefly review the research activities that
were triggered by identification for control, as well as some other topics that have
seen important developments in the last decade.

6.3

Quantification of model uncertainty

The demands of robust control theory for adequate uncertainty sets triggered a
lot of interest for the estimation of uncertainty sets from data. It is fair to say
that most of the robust control theory developed in the eighties had been based
on a priori assumed uncertainty bounds on model errors and on the noise. In

16

System Identification without Lennart Ljung: what would have been different?

the context of system identification, the estimation of model errors from data is
of course of interest in its own right; indeed a reputable engineer should never
deliver a product (a model, in this case) without a statement about its error margins. In this authorss opinion, this activity on estimation of uncertainty sets was
erroneously put under the umbrella of identification for control; indeed, in most
of the work on data-based estimation of uncertainty sets of the last decade, the
control objective is not taken into account in the identification design.
A wide range of new techniques were developed to provide error bounds on
estimated models, in all kinds of shapes and frameworks, using time domain,
frequency domain, H , l1 , probabilistic, worst case, set membership and other
methods: see Helmicki et al. (1991); Goodwin et al. (1992); Poolla et al. (1994);
Smith and Dahleh (1994); de Vries and Van den Hof (1995); Kosut (1995);
Hakvoort and Van den Hof (1997); Giarr et al. (1997). Even though it is now a
few years old, probably one of the best presentations of these different methods
is the survey paper Ninness and Goodwin (1995).

6.4

Closed-loop identification revisited

For some reason, the work of the seventies on closed-loop identification had
stopped at the question of identifiability, i.e. under what conditions do the parameter estimates converge to the true parameters in the case where the system
is in the model set (S M). The influence of the experimental conditions on
bias error distribution in the case of restricted complexity models, as well as on
asymptotic variance, had not been addressed.
One of the important lessons that emerged from the study of the interplay between identification and control is the benefit of closed-loop identification when
the model is to be used for control design. Until the late eighties, it was commonly accepted that closed-loop identification was preferably to be avoided. In
identification for control with reduced order models, the required connection between the control performance criterion (obviously a closed-loop criterion) and
the identification criterion, established the need for closed-loop identification,
as described above. In the ideal context of optimal experiment design with full
order models, optimality of closed-loop identification was actually established
when the model is to be used for control design with a noise rejection objective:
Gevers and Ljung (1986); Hjalmarsson et al. (1996); Forssell and Ljung (2000).
This observation triggered an important new activity in the design of special
purpose closed-loop identification methods, the main goal pursued by these new
methods being to obtain a better handle on the bias error in closed-loop identification: Hansen et al. (1989); Van den Hof and Schrama (1993); Van den Hof
et al. (1995); Forssell and Ljung (1999).

Michel Gevers

6.5

17

Optimal experiment design for control

In the 1970s, optimal input design for system identification was an active area
of research, with different quality measures being used for this optimal design:
Mehra (1974); Zarrop (1979); Goodwin and Payne (1977). The questions at that
time addressed open-loop identification and focused on quality measures of the
parameter covariance matrix P . That work ground to a halt for about fifteen
years. The new paradigm of identification as a design problem has given it a
new lease of life. The recent work on the connection between model uncertainty
sets obtained by identification and corresponding sets of robust controllers have
clearly put this subject in the limelight again: Gevers et al. (1999), Forssell and
Ljung (2000), Hildebrand and Gevers (2003); Lindqvist and Hjalmarsson (2001);
Barenthin et al. (2005); Bombois et al. (2004). The new emphasis is to establish
a direct link between the experimental conditions under which a model is identified (together with its uncertainty set) and the performance of the controller that
results from the use of the model and its estimated uncertainty set.

6.6

Frequency domain identification

Another area of important activity in the nineties has been frequency domain
identification. In the identification community, there has historically been no
more than a polite attention paid to frequency domain identification. Every now
and then, some of the experts in the area published a paper whose main aim was
to reassure the community with a message that sounded essentially like yes,
indeed, we can also handle frequency domain data: Ljung and Glover (1981).
Things changed drastically at the end of the eighties, with a convergence of
efforts arising from two completely different horizons.
During the eighties, the robust control community had developed most of
its analysis and design tools in the frequency domain; thus, there was a
great demand for tools that would enable one to obtain frequency domain
models and, even more importantly, frequency domain uncertainty descriptions from data. This led, at the end of the eighties, to the development of
a number of new interpolation techniques that were using noisy pointwise
frequency domain transfer function measurements as their data Parker and
Bitmead (1987); Partington (1991); Gu and Khargonekar (1992);
At about the same time, Pintelon and Schoukens had independently developed frequency domain identification techniques, essentially based on
the Maximum Likelihood principle. With their instrumentation and measurement background, they were interested in methods that would deliver
reliable models for devices under test, through the application of short input data sequences Pintelon and Schoukens (1990).

18

System Identification without Lennart Ljung: what would have been different?

Aware of the interest for frequency domain identification emanating from

the robust control community, Schoukens and Pintelon continued their work on
frequency domain identification all through the nineties, with specific and important contributions on the use of periodic excitation and of maximally informative
input signals. Their book Pintelon and Schoukens (2001) provides a comprehensive treatment of frequency domain identification. In Ljung (2006a), Ljung
updates the comparative analysis of Ljung and Glover (1981) between time and
frequency domain identification on the basis of the new understanding gained
about frequency domain identification over the last twenty years.

6.7

Some other areas of activity in the last decade

The areas mentioned above have been areas of major activity in the nineties. With
the exception of subspace-based identification, most of these areas have had their
origin in the new engineering view of identification as a design problem. This has
created a lot of interest in all aspects of identification for control, which turned
out to require better methods for the quantification of uncertainty, a better understanding of the bias issues in closed-loop identification and the development of
frequency domain identification methods that were as much as possible compatible with frequency domain based robust control analysis tools. But these were
by no means the only areas of activity in system identification.
Interesting progress has been made on the use of alternative basis functions
(other than the shift operator) for the representation of input-output models, such
as Laguerre, Kautz, and other generalized orthonormal basis functions. Such
alternative bases can not only lead to more compact descriptions when some
prior knowledge is available about the system, but they have also led to improved
formulas for the estimation of the variance of black box transfer function models:
Heuberger et al. (1995); Ninness et al. (1999a,b).
Finally, this overview of recent activity would not be complete without a mention of the topic of identification of nonlinear systems. As is very nicely stated
in Deistler (2002), identification of nonlinear systems is like a statement about
non-elephant zoology. Indeed, the area is vast, and the approaches are many,
ranging from semi-parametric identification using, e.g, neural network methods,
to the development of ad hoc techniques for specific classes of nonlinear systems such as, e.g. compartmental models. The paper Sjberg et al. (1995), even
though no longer recent, is still a very valuable overview and introduction to the
subject.

Michel Gevers

19

So what would have been different?

What would have been different without Lennart Ljung? Clearly, it is impossible to guess the direction that system identification would have taken without
Lennart Ljung. Somebody else might have had the same ideas at about the same
time; who knows? But clearly, Ljung is responsible for some key and judicious
trajectory changes in the development of system identification theory, which have
profoundly altered our ways of thinking about the topic. If I were to single out
what I view as his most important contributions, I would list the following.
His introduction of a solid identification framework with a clear separation
between the parametric model, seen as a way of computing predictions,
and a criterion, whose minimization yields an estimated model.
His move from properties of parameter estimates to properties of transfer
function estimates, with the introduction of the concept of bias and variance of transfer function estimates. A critical feature here was the introduction of , Ljungs baby.
The formulation of system identification as a design problem, with the
introduction of the objective as a weighting in the identification criterion.
The parallel development of the theory and of the identification toolbox.
The impact of this effort of Ljung on the spread of system identification in
academia and industry has been immense.

Concluding remarks

I have presented my personal views on the development of system identification

in the control community, as I have observed it over a period of thirty years,
both as a student of the subject eager to learn and understand the work of my
colleagues, and as an active participant in these developments. To paraphrase
Ljung (2006b), I shall say This chapter is not a survey": see Magritte (1929).
Rather it should be read as a story told with the eyes and the prejudices of one
of the actors in the field. I tend to believe that the way a particular field of
science develops depends on a combination of two forces: the socio-technical
environment created by the evolution of the neighbouring fields of science and
by the demands of the applications world, and the creative role played by a few
individuals who suddenly make it possible to venture into a totally new direction
or to establish a useful link with another field of science that sheds new light on
the subject. Clearly, if there is one person who has produced these creative moves
in system identification over the last 30 years, it is Lennart Ljung.

20

System Identification without Lennart Ljung: what would have been different?

Acknowledgments
This chapter presents research results of the Belgian Programme on Interuniversity Poles of Attraction, initiated by the Belgian State, Prime Ministers Office
for Science, Technology and Culture. The scientific responsibility rests with its
author.

References
Akaike, H. (1974). Stochastic theory of minimal realization, IEEE Trans. Automatic Control 26: 667673.
Akaike, H. (1975). Markovian representation of stochastic processes by canonical variables, SIAM Journal on Control and Optimization 13: 162173.
Anderson, B. and Gevers, M. (1982). Identifiability of linear stochastic systems
operating under linear feedback, Automatica, 18 2: 195213.
Anderson, B., Moore, J. and Hawkes, R. (1978). Model approximation via prediction error identification, Automatica 14: 615622.
strm, K. (1980). Maximum likelihood and prediction error methods, Automatica 16: 551574.
strm, K. and Bohlin, T. (1965). Numerical identification of linear dynamic
systems from normal operating records, Proc. IFAC Symposium on SelfAdaptive Systems, Teddington, UK.
strm, K. and Eykhoff, P. (1971). System identification A survey, Automatica
7: 123162.
strm, K. and Nilsson, J. (1994). Analysis of a scheme for iterated identification and control, Proc. IFAC Symp. on Identification, Copenhagen, Denmark,
pp. 171176.
Barenthin, M., Jansson, H. and Hjalmarsson, H. (2005). Applications of mixed
H2 and H input design in identification, 16th IFAC World Congress on Automatic Control, paper 03882, Prague.
Bayard, D., Yam, Y. and Mettler, E. (1992). A criterion for joint optimization of
identification and robust control, IEEE Trans. Automatic Control 37: 986991.
Bombois, X., Gevers, M. and Scorletti, G. (2000). A measure of robust stability
for an identified set of parametrized transfer functions, IEEE Transactions on
Automatic Control 45(11): 21412145.

Michel Gevers

21

Bombois, X., Scorletti, G., Gevers, M., Hildebrand, R. and Van den Hof, P.
(2004). Cheapest open-loop identification for control, CD-ROM Proc. 33rd
IEEE Conf on Decision and Control, The Bahamas, pp. 382387.
Box, G. and Jenkins, G. (1970). Time Series Analysis, forecasting and control,
Holden-Day, Oakland, California.
Chiuso, A. and Picci, G. (2002). Geometry of oblique splitting subspaces, minimality and hankel operators, Directions in Mathematical Systems Theory and
Optimization, A. Rantzer and C. Byrnes Eds, Springer-Verlag, New York.
Chui, N. and Maciejowski, J. (1996). Realization of stable models with subspace
methods, Automatica 32(11): 15871595.
Clark, J. (1976). The consistent selection of parametrizations in systems identification, Proc. Joint Automatic Control Conference, Purdue University.
de Callafon, R. and Van den Hof, P. (1997). Suboptimal feedback control by a
scheme of iterative identification and control design, Mathematical Modelling
of Systems 3(1): 77101.
de Vries, D. and Van den Hof, P. (1995). Quantification of uncertainty in transfer
function estimation: a mixed probabilistic - worst-case approach, Automatica
31: 543558.
Deistler, M. (2002). System identification and time series analysis: Past, present,
and future, Stochastic Theory and Control, Festschrift for Tyrone Duncan,
Springer, Kansas, USA, pp. 97108.
Deistler, M. and Gevers, M. (1981). Some properties of the parametrization of
arma systems with unknown order, Multivariate Analysis 11: 474484.
Faurre, P. (1976). Stochastic realization algorithms, System identification: Advances and Case Studies R.K. Mehra and D.G. Lainiotis Eds., Academic
Press, New York.
Forssell, U. and Ljung, L. (1999). Closed-loop identification revisited, Automatica 35: 12151241.
Forssell, U. and Ljung, L. (2000). Some results on optimal experiment design,
Automatica 36: 749756.
Gauss, C. (1809). Teoria Motus Corporum Coelestium in Sectionibus Conicis
Solem Ambientium, Reprinted translation: Theory of Motion of the Heavenly
Bodies Moving about the Sun in Conic Sections, Dover, New York.

22

System Identification without Lennart Ljung: what would have been different?

Gevers, M. (1991). Connecting identification and robust control: A new challenge, Proc. IFAC/IFORS Symposium on Identification and System Parameter
Estimation, Budapest, Hungary, pp. 110.
Gevers, M. (2005). Identification for control: from the early achievements to the
revival of experiment design, European Journal of Control 11: 118.
Gevers, M., Bombois, X., Codrons, B., Bruyne, F. D. and Scorletti, G. (1999).
The role of experimental conditions in model validation for control, in
A. Garulli, A. Tesi and A. Vicino (eds), Robustness in Identification and Control - Proc. of Siena Workshop, July 1998, Vol. 245 of Lecture Notes in Control
and Information Sciences, Springer-Verlag, pp. 7286.
Gevers, M., Bombois, X., Codrons, B., Scorletti, G. and Anderson, B. (2003).
Model validation for control and controller validation in a prediction error
identification framework - Part I: theory, Automatica 39(3): 403415.
Gevers, M. and Kailath, T. (1973). An innovations approach to least-squares
estimation, part vi: Discrete-time innovations representations and recursive
estimation, IEEE Transactions Automatic Control 18: 588600.
Gevers, M. and Ljung, L. (1986). Optimal experiment designs with respect to the
intended model application, Automatica 22: 543554.
Giarr, L., Milanese, M. and Taragna, M. (1997). H identification and model
quality evaluation, IEEE Trans. Automatic Control 42(2): 188199.
Goodwin, G., Gevers, M. and Ninness, B. (1992). Quantifying the error in
estimated transfer functions with application to model order selection, IEEE
Trans. Automatic Control 37: 913928.
Goodwin, G. and Payne, R. (1977). Dynamic System Identification: Experiment
Design and Data Analysis, Academic Press, New York.
Gu, G. and Khargonekar, P. (1992). A class of algorithms for identification in
H , Automatica 28: 299312.
Gustavsson, I., Ljung, L. and Sderstrm, T. (1977). Identification of processes
in closed loop - identifiability and accuracy aspects, Automatica 13: 5975.
Hakvoort, R. and Van den Hof, P. (1997). Identification of probabilistic system
uncertainty regions by explicit evaluation of bias and variance errors, IEEE
Trans. Automatic Control 42(11): 15161528.
Hansen, F., Franklin, G. and Kosut, R. (1989). Closed-loop identification via the
fractional representation: Experiment design, Proc. American Control Conference pp. 14221427.

Michel Gevers

23

Hazewinkel, M. and Kalman, R. (1976). On invariants, canonical forms and moduli for linear constant, finite-dimensional dynamical systems, Lecture Notes
in Economics and Mathematical Systems, Vol. 131, Springer-Verlag, Berlin,
pp. 4860.
Helmicki, A., Jacobson, C. and Nett, C. (1991). Control oriented system identification: A worst-case/deterministic approach in H , IEEE Trans. Automatic
Control AC-36: 11631176.
Heuberger, P., Bosgra, O. and Van den Hof, P. (1995). A generalized orthonormal basis for linear dynamical systems, IEEE Trans. Automatic Control AC40(3): 451465.
Hildebrand, R. and Gevers, M. (2003). Identification for control: optimal input design with respect to a worst-case -gap cost function, SIAM Journal on
Control and Optimization 41(5): 15861608.
Hjalmarsson, H. (2005). From experiment design to closed-loop control, Automatica 41: 393438.
Hjalmarsson, H., Gevers, M. and Bruyne, F. D. (1996). For model-based control design, closed-loop identification gives better performance, Automatica
32: 16591673.
Ho, B. and Kalman, R. (1965). Effective construction of linear state-variable
models from input-output functions, Regelungstechnik 12: 545548.
Katayama, T. and Picci, G. (1999). Realization of stochastic systems with exogenous inputs and subspace identification methods, Automatica 35: 16351652.
Kosut, R. (1995). Uncertainty model unfalsification : a sytem identification
paradigm compatible with robust control design, Proc. 34th IEEE Conference
on Decision and Control, New Orleans,LA.
Kosut, R. and Anderson, B. (1994). Least-squares parameter set estimation for robust control design, Proc. American Control Conference, Baltimore,Maryland, pp. 30023006.
Larimore, W. (1990). Canonical variate analysis in identification, filtering, and
adaptive control, Proc. 29th IEEE Conf. on Decision and Control, Honolulu,
Hawaii, pp. 596604.
Lee, W., Anderson, B., Kosut, R. and Mareels, I. (1993). A new approach to
adaptive robust control, Int. Journal of Adaptive Control and Signal Processing
7: 183211.

24

System Identification without Lennart Ljung: what would have been different?

Lindqvist, K. and Hjalmarsson, H. (2001). Identification for control: adaptive

input design using convex optimization, CD-ROM Porc. of 40th IEEE Conf.
on Decision and Control, Orlando, Florida.
Liu, K. and Skelton, R. (1990). Closed loop identification and iterative controller
design, 29th IEEE Conf on Decision and Control pp. 482487.
Ljung, L. (1976). On consistency and identifiability, Mathematical Programming
Study 5: 169190.
Ljung, L. (1978). Convergence analysis of parametric identification methods,
IEEE Trans. Automatic Control AC-23: 770783.
Ljung, L. (1985). Asymptotic variance expressions for identified black-box transfer function models, IEEE Trans. Automatic Control AC-30: 834844.
Ljung, L. (1987). System Identification: Theory for the User, Prentice-Hall,
Englewood Cliffs, NJ.
Ljung, L. (2006a). Frequency domain versus time domain methods in system
identification, in B. Francis and J. Willems (eds), Control of Uncertain Systems: Modelling, Approximation and Design, Springer Verlag, LNCIS Series,
pp. 277291.
Ljung, L. (2006b). Some aspects on nonlinear system identification, CD-Rom
Proc. of 14th IFAC Symposium on System Identification, Newcastle, Australia,
pp. 110121.
Ljung, L. and Caines, P. (1979). Asymptotic normality of prediction error estimators for approximative system models, Stochastics 3: 2946.
Ljung, L. and Glover, K. (1981). Frequency domain versus time domain methods
in system identification, Automatica 17(1): 7186.
Magritte, R. (1929). Ceci nest pas une pipe, Los Angeles County Museum of Art,
Los Angeles, USA.
Mkil, P., Partington, J. and Gustafsson, T. (1995). Worst-case control-relevant
identification, Automatica 31(12): 17991819.
Mehra, R. (1974). Optimal input signals for parameter estimation in dynamic
systems - survey and new results, IEEE Trans. on Automatic Control AC19(6): 753768.
Ng, T., Goodwin, G. and Anderson, B. (1977). Identifiability of mimo linear
dynamic systems operating in closed loop, Automatica 13: 477485.

Michel Gevers

25

Ninness, B. and Goodwin, G. (1995). Estimation of model quality, Automatica

31(12): 3274.
Ninness, B., Hjalmarsson, H. and Gustafsson, F. (1999a). Asymptotic variance
expressions for output error model structures, 14th IFAC World Congress, Beijing, P.R. China.
Ninness, B., Hjalmarsson, H. and Gustafsson, F. (1999b). On the fundamental role of orthonormal bases in system identification, IEEE Transactions on
Automatic Control 44(7): 13841407.
Parker, P. and Bitmead, R. (1987). Adaptive frequency response identification,
Proc. CDC .
Partington, J. (1991). Robust identification and interpolation in H , International Journal of Control 54: 12811290.
Picci, G. (1982). Some numerical aspects of multivariable systems identification,
Math. Programming Studies 18: 76101.
Pintelon, R. and Schoukens, J. (1990). Robust identification of transfer functions
in the s- and z-domains, IEEE Trans. on Instrumentation and Measurement
39(4): 565573.
Pintelon, R. and Schoukens, J. (2001). System Identification - A Frequency Domain Approach, IEEE Press, New York.
Poolla, K., Khargonekar, P., Tikku, A., Krause, J. and Nagpal, K. (1994). A
time-domain approach to model validation, IEEE Trans. Automatic Control
39: 951959.
Rissanen, J. (1982). Estimation of structure by minimal description length, Circuits, Systems, and Signal Processing 1: 395406.
Schrama, R. (1992). Accurate identification for control: The necessity of an
iterative scheme, IEEE Trans. on Automatic Control 37: 991994.
Schrama, R. and Bosgra, O. (1993). Adaptive performance enhancement by
iterative identification and control design, Int. Journal of Adaptive Control and
Signal Processing 7(5): 475487.
Sin, K. and Goodwin, G. (1980). Checkable conditions for identifiability of linear
systems operating in closed loop, IEEE Trans. Automatic Control AC-25: 722.
Sjberg, J., Zhang, Q., Ljung, L., Benveniste, A., Deylon, B., Glorennec, P.-Y.,
Hjalmarsson, H. and Juditsky, A. (1995). Nonlinear black-box modeling in
system identification: a unified overview, Automatica 31(12): 16911724.

26

System Identification without Lennart Ljung: what would have been different?

Smith, R. S. and Dahleh, M. (1994). The Modeling of Uncertainty in Control

Systems, Lecture Notes in Control and Information Sciences, Volume 192,
Springer-Verlag, London.
Sderstrm, T., Ljung, L. and Gustavsson, I. (1976). Identifiability conditions for
linear multivariable systems operating under feedback, IEEE Trans. Automatic
Control AC-21: 837840.
Sderstrm, T. and Stoica, P. (1989). System Identification, Prentice-Hall International, Hemel Hempstead, Hertfordshire.
Van den Hof, P. and Schrama, R. (1993). An indirect method for transfer function
estimation from closed loop data, Automatica 29: 15231527.
Van den Hof, P., Schrama, R., de Callafon, R. and Bosgra, O. (1995). Identification of normalized coprime plant factors from closed-loop experimental data,
European Journal of Control 1: 6274.
Van Overbeek, A. and Ljung, L. (1982). On-line structure selection for multivariable state-space models, Automatica 18: 529544.
Van Overschee, P. and De Moor, B. (1994). N4SID: Subspace algorithms for
the identification of combined deterministic-stochastic systems, Automatica
30(1): 7593.
Verhaegen, M. (1994). Identification of the deterministic part of mimo state
space models given in innovations form from input-output data, Automatica
30(1): 6174.
Viberg, M. (1995). Subspace-based methods for the identification of linear timeinvariant systems, Automatica 31(12): 18351851.
Wahlberg, B. and Ljung, L. (1986). Design variables for bias distribution in
transfer function estimation, IEEE Trans. Automatic Control AC-31: 134144.
Wertz, V., Gevers, M. and Hannan, E. (1982). The determination of optimum
structures for the state-space representation of multivariable stochastic processes, IEEE Transactions on Automatic Control 27(6): 12001211.
Zang, Z., Bitmead, R. and Gevers, M. (1995). Iterative weighted least-squares
identification and weighted LQG control design, Automatica 31(11): 1577
1594.
Zarrop, M. (1979). Optimal experiment design for dynamic system identification,
Lecture Notes in Control and Information Sciences, Vol. 21, Springer Verlag,
Berlin, New York.